OpenAI’s latest AI models have a new safeguard to prevent biorisks

Open says it has deployed a new system to monitor its latest AI rational models, O3 and O4-MiniFor requests related to biological and chemical threats. The goal of the system is to prevent models from giving suggestions that may instruct someone to carry out a potentially harmful attack, According to the OPENAIThe

The agency says that the 3 and the 4-Mint AAII represents a meaningful capacity compared to the previous models of the OpenAI, and thus create new risks in the hands of bad actors. According to the internal criteria of the Openai, 3, especially in the way of creating a certain type of biological threats, is more efficient in answering questions. This reasons-and to reduce other risks-Opnai has created a new monitoring system, describing the company as a “protection-centric argument monitor”.

The contents of the OPENY are rational custom-educated monitors about the policies and 3 and the 4-min-min. It is designed to detect prompts related to biological and chemical risk and to indicate models to refuse to advice on those topics.

For setting up a baseline, Red Timers and 3 and 4-Mins in OpenAE spent about 1,000 hours flagging the Biorisk conversation from 4-min. According to the Openai, during an experiment that imitates the “blocking logic” of its protection monitor, models refused to respond to 98.7% of the time, according to Openai.

Opena acknowledges that its test has not been an account for people who can try new prompts after the monitor is blocked, which is why the company says it will depend on human observation.

According to the agency, the “high risk” of the 3 and 4-Minit Openai does not exceed the threshold. However, the OpenAI, compared to the O 1 and GPT -4, says that the primary versions of 3 and 4 -min -minimum proved to be more helpful in answering the question surrounding biological weapons development.

Openai’s recently updated update is actively tracking how its models can make its models easier to develop chemical and biological threats to corrupt users Preparation structureThe

OpenAI depends on increasingly automatic systems to reduce risk from its models. For example, to prevent GPT -4O’s native image generator from creating elements of child sexual abuse (CSAM) (Open is saying that it uses an argument similar to the company deployed for 3 and 4-min.

Nevertheless, several researchers have raised concerns that Open is not giving priority to security as much as he should. One of the company’s Red-Timing Partners, Mater, says that there was a relatively short time to test a standard and 3 for fraudulent behavior. Meanwhile, Openai decided not to release a Its GPT -5.1 model is the protection report for the modelWhich was launched early this week.